Partial Coherence for Object Recognition and Depth Sensing

TL;DR

We address how illumination coherence, parameterized by the transverse coherence length $l_c$, affects object recognition and depth sensing in simulated imaging. The approach combines computational partial coherence via dynamic random phase screens with angular spectrum propagation and a ResNet-18 classifier trained on MNIST and Fashion-MNIST, evaluated in direct and diffuser-scattered scenes, with coherence varied across trials. Results show that increasing $l_c$ augments image information as measured by two-dimensional information entropy $H$, and correspondingly improves recognition and depth-sensing accuracy, with saturation at higher coherence; the diffuser reduces absolute accuracy but preserves the monotonic trend and entropy behavior. The work highlights that high—but not strictly perfect—coherence can suffice for robust CV performance, offering guidance for illumination design in practical systems and motivating extensions to complex imaging contexts.

Abstract

We show a monotonic relationship between performances of various computer vision tasks versus degrees of coherence of illumination. We simulate partially coherent illumination using computational methods, propagate the lightwave to form images, and subsequently employ a deep neural network to perform object recognition and depth sensing tasks. In each controlled experiment, we discover that, increased coherent length leads to improved image entropy, as well as enhanced object recognition and depth sensing performance.

Figures (6)

• Figure 1: Direct imaging scenario experimental setup and results. (a) Schematic diagram of basic scene setup. The coherent light emitted by the source becomes partially coherent after decoherence. (b) Example images of imaging results of different objects under varying degrees of coherence. The images with a low $l_{c}$ value are also quite dark, so we have normalized the images to improve their visibility. (c) Object recognition accuracy and two-dimensional information entropy curves.
• Figure 2: The degree of coherence measurement schematic and results. (a) Schematic diagram of measurement optical path. (b) Quantitative relationship between $l_{c}$ and degree of coherence. (c) Interference fringe images at different degrees of coherence.
• Figure 3: The neural network architecture utilized for the image classification. It consists of a combination of convolutional and fully connected layers. The abbreviation "Conv" stands for a convolutional layer, "Avg" represents an average pooling layer, and "Fc" denotes a fully connected layer.
• Figure 4: Scattering imaging scenario experimental setup and results. (a) Schematic diagram of the scenario after the addition of a diffuser. (b) Image examples of imaging results of different objects with different coherence degrees. (c) Accuracy curves and two-dimensional information entropy curves with and without diffuser.
• Figure 5: Examples and results of imaging of Fashion-MNIST object set at different coherence levels. (a) Example of imaging for the scenario depicted in Figure \ref{['fig:1']}(a). (b) Example of imaging for the scenario depicted in Figure \ref{['fig:4']}(a). (c) Accuracy curves and two-dimensional information entropy curves with and without diffuser.